Fix payload model name when model id is a URL (#2911)

* fix default model name when model id is a URL

* better

* Update test

Co-authored-by: Lucain <[email protected]>

---------

Co-authored-by: Lucain <[email protected]>

Author: hanouticelina

src/huggingface_hub/inference/_providers/hf_inference.py

@@ -84,7 +84,11 @@ def __init__(self):
         super().__init__("text-generation")
 
     def _prepare_payload_as_dict(self, inputs: Any, parameters: Dict, mapped_model: str) -> Optional[Dict]:
-        payload_model = "tgi" if mapped_model.startswith(("http://", "https://")) else mapped_model
+        payload_model = parameters.get("model") or mapped_model
+
+        if payload_model is None or payload_model.startswith(("http://", "https://")):
+            payload_model = "dummy"
+
         return {**filter_none(parameters), "model": payload_model, "messages": inputs}
 
     def _prepare_url(self, api_key: str, mapped_model: str) -> str:

tests/test_inference_providers.py

@@ -305,6 +305,53 @@ def test_prepare_request_conversational(self):
             "messages": [{"role": "user", "content": "dummy text input"}],
         }
 
+    @pytest.mark.parametrize(
+        "mapped_model,parameters,expected_model",
+        [
+            (
+                "username/repo_name",
+                {},
+                "username/repo_name",
+            ),
+            # URL endpoint with model in parameters - use model from parameters
+            (
+                "http://localhost:8000/v1/chat/completions",
+                {"model": "username/repo_name"},
+                "username/repo_name",
+            ),
+            # URL endpoint without model - fallback to dummy
+            (
+                "http://localhost:8000/v1/chat/completions",
+                {},
+                "dummy",
+            ),
+            # HTTPS endpoint with model in parameters
+            (
+                "https://api.example.com/v1/chat/completions",
+                {"model": "username/repo_name"},
+                "username/repo_name",
+            ),
+            # URL endpoint with other parameters - should still use dummy
+            (
+                "http://localhost:8000/v1/chat/completions",
+                {"temperature": 0.7, "max_tokens": 100},
+                "dummy",
+            ),
+        ],
+    )
+    def test_prepare_payload_as_dict_conversational(self, mapped_model, parameters, expected_model):
+        helper = HFInferenceConversational()
+        messages = [{"role": "user", "content": "Hello!"}]
+
+        payload = helper._prepare_payload_as_dict(
+            inputs=messages,
+            parameters=parameters,
+            mapped_model=mapped_model,
+        )
+
+        assert payload["model"] == expected_model
+        assert payload["messages"] == messages
+
 
 class TestHyperbolicProvider:
     def test_prepare_route(self):